Slashdot Mirror
Nanotechnology And The Law of Accelerating Returns
digitect writes: "The article More More More at Reason is a good overview of the increasing rate of acceleration for technology. It includes references to nanotube technology, nanobots and estimations of gross computing power in the near and far future.
Frankly, I doubt we will ever develop computers with the sophisticated power of even a mouse brain, although many may protest that we already have exceeded their gross power. I believe that things like perception and reasoning are beyond the scope of raw power. But it's a fun read anyway."
Typically, the "computers will be more powerful than humans!" comments come from newbies, while "computers will get faster but will be distinctly different than brains" is what you hear from people who have been involved in AI for a while.
The tremendous speed increases in computing hardware are often mistaken for something deeper. We're writing larger applications, yes, but they're not necessarily more stable or more advanced in a way that's different than simply adding more features. If anything, we're starting to come to the realization that simpler is better, or at least that having straightforward goals is much better than shooting for extremes.
Take compilers, for example. In the 1970s, two top goals of compiler writers were "incredibly high levels of optimization" and "automatic correction of user errors." Today the goal is more conservative: "go for a straightforward implementation that will have the fewest problems." It isn't worth doing over-the-top optimization if you're trading a 0.5% speed increase for greatly increased code complexity. As a result, more compiler writers have taken a conservative approach. In terms of correcting user errors, it is simpler and more predictable to simply report errors as they are found. Trying to be smart causes more trouble than it is worth ("How can my program be wrong if it compiles and runs?").
Complexity is a limiting factor in grandoise plans for AI.
This stuff sounds a lot like the science fiction of the forties and fifties..."By 1980, people will travel to their offices in automated helicopters! People will fly to the moon and mars on a regular basis!" Many science fiction novels of the time portrayed extravagant space-travel, and yet had humans doing the navigational computations on board these spacecraft, completely missing the coming dominance of electronic computers in computations. The popular imagination of future technology was simply an extrapolation of the current technology - "Fast airplanes, bigger rockets, atomic power!" - when in reality the technology of the future has turned out to be things that people could not have imagined at the time (the world-wide-web, biotechnology, fax machines, massive computational power...) In the same vein, the breakthrough technology of the next 100 years will most likely consist of things that haven't even occured to us yet -- not simply faster and faster computers, smaller and smaller robots, better and better bioengineering, but rather something else entirely.
All is Number -Pythagoras.
I refer, of course, to the human brain. All it is, is an organic, massively-parallel processing environment with an equally massive and complex boot-rom, and a uniquely flexible OS.
So, why is it ok to have a jellyware CPU system with intuition, and not a hardware one ?? The difference is likely to be a matter of "hardware" complexity, and the "software" running on it. . .
Frankly, I doubt we will ever develop computers with the sophisticated power of even a mouse brain, although many may protest that we already have exceeded their gross power. I believe that things like perception and reasoning are beyond the scope of raw power.
Just to offer my viewpoint... The brain is slow, but massively parallel and interconnected in a vast array of various neural networks. Inherently, the brain is analog -- down to the quanta of electrons involved in the chemical reactions.
In order to simulate that in a computer that executes things very quickly, but serially, would require a HUGE AMOUNT of computing power. You'd have to be able to simulate time-slices as small as those significant in the brain.
However, if we were to take several slow processors, and network them together in parallel, we'd probbably get a lot closer for a lot less.
I don't believe consciencenous is anything special. Its just the superposition of hundreds or thousands of neural networks all owrking together. Heck, at one time, man-kind thought the motion of the planets and stars were just too complicated to ever figured out, so they were labeled as something mysterious and never to be known. We shouldn't make that same mistake with the brain and mind simple because it appears at present to be too complicated to figure out.
I believe that things like perception and reasoning are beyond the scope of raw power."
Actually, these two areas of artificial intelligence that are probably understood better than any others. (Language and Memory--now those are problems people are still clueless about, IYAM.) Neuroscientists have mapped out the perceptual system with great detail (at least the visual perceptual system), and there are some fairly advanced neural network models that embody these findings. On the other hand, Newell and Simon were able to understand and explain many kinds of "reasoning" very well in the 1970s--today the main descendent of this work, "SOAR", can work with 10,000 or more rules. It can fly planes in simulated combat and make strategic and tactical decisions. Maybe it is unable to do everything a pilot does, but I would argue that it is still reasoning.
So, it is technically correct to say that these things are beyond the scope of raw power, but the theoretical advancements have already been made. The only thing holding these system back from real-time performance is raw power.
Processors provide the "parts" of computation by physically performing the actual instructions used. These computers basically allow numerical operations, memory access, and branching. That doesn't seem like much, but it's "Turing complete," which means that (if you buy the Turing hypothesis) everything which is computable can be computed with such instructions. We have all the parts of computation we need, and they're getting faster all the time.
But the software still lags. We have "computationally intense" software, but that's not the same as complex software. 3D games always push the envelope of computer capability because just when you think you've got enough computing power, id throws more triangles and more textures at the problem. That's a quantitative change, but not a qualitative change.
When we look at all of the other software produced, it seems that if the software is marginally complex (think of your favorite program here), it's buggy as hell. Reducing the bugs in the software requires more effort; an exponential amount of effort as the complexity increases.
That's why we've seen the speed of computer hardware shoot through the roof, and the complexity of computer software plod along, unable to keep up. Producing complex software is an NP-complete problem. (/me ducks the flames of the math people in the audience.)
If you'll permit me to play pundit for a second: I think we'll reach these so-called "milestones" that the AI people and the nanotech people keep giving us and realize that while we can manufacture a computer with the MIPS/FLOPS/whatever of a mouse/dog/human brain, we don't have the slightest idea how to string all of that power together to actually perform the operations of the mouse/dog/human brain.
Your computer will get 10,000 fps with 6e10 textured polygons in Quake XXXVI, but it still won't be able to learn a new language.
Do we have to understand it, first? It seems like if we could cheat a bit, and just model a mouse brain inside a physics simulation, we'd have a computer engaging in perception-like tasks. The obvious drawback to this is that it wouldn't be capable of doing anything a mouse couldn't do (and would lead to a flurry of /. posts along the lines of "Imagine a Beowulf cluster of these things! They could run through mazes and find cheese! The possibilities are limitless!"). However, by modelling a mouse brain, scientists would be able to better "fiddle" with it and understand it, possibly leading to a more practical understanding of perception.
And to get further off on a tangent (but hopefully remaining within the realm of a worthwhile discussion), we suddenly open up a whole can of worms with regard to creating a machine-based consciousness. In my own opinion (and this is just opinion here), a hypothetically powerful/complete enough simulation of a human brain approaches consciousness. I'm of the opinion that an actual living, breathing human is just such a simulation via chemical means. I'm a little afraid of the ethical consequences when we gain the ability to create neural networks with complexity rivaling that of our own brains; one could argue that it's of even greater ethical concern than human cloning. A human clone, at least, has the benefit of being inarguably human (barring something really weird like a gorilla/human hybrid), and would thus be protected by normal laws.
And then we could build a beowulf cluster of these!
Goodbye Karma.
A large, networked system of analog neurons might just do the trick here as for creating a system with the intellegence of a mouse. But absent any good way to deliver, register, and respond to stimulus, this would be one crazy machine. It simply wouldn't have enough information to act, any way to deal with information sent to it, or any way to figure out whether its actions were appropriate or inappropriate (it would need a complex systems of rewards and punishments and some sort of inherant internal mapping of neurons to stimuli and responses.) To wit, experiments have shown that if you cluster a bunch of analog neurons together, it will think random thoughts until you bother to shut it off.
Plus it would need to "eat", self-repair, purge unneeded inputs (both by discarding unsupported hypothesises [Is this a cat? It does not look like a cat. It is not a cat.] and if it eats it will have to poop), and eventually defend itself against hazards. In other words, mice will be "better" for a long long time.
-Ben
"Ancillary does not mean you get to rule the world." --U.S. Circuit Judge Harry Edwards, speaking to the FCC's lawyer
Frankly, I doubt we will ever develop computers with the sophisticated power of even a mouse brain ...
And 640k should be enough for anybody, right? :-)
All this would be done automatically, effortlessly, without human hands or labor, by a fleet of tiny, invisible robots
Thats funny, thats exactly how my boss thinks work gets done too.
http://twitter.com/onion2k